Source-linked AI summary
Controlled-Minimax: Stabilizing Adversarial Optimization via Control Theory
Tianjiao Luo, Jianfei Chen, Jun Zhu
TL;DR
Minimax training can oscillate or diverge, and existing guarantees are limited in the non-convex, non-concave settings common in deep learning. Controlled-Minimax models training as a second-order ODE and introduces feedback control, with local stability guarantees and empirical suppression of oscillations across synthetic benchmarks and StyleGAN2-ada.
Problem
In non-convex, non-concave deep-learning settings, Gradient Descent-Ascent may fail to reach a stable Nash equilibrium, while existing guarantees and stabilization approaches have limitations.
Source: §1 · pp. 1–2Method
Controlled-Minimax models minimax training as a second-order ODE and injects a designed feedback controller into the dynamics.
Source: §1 · p. 2Results
Takeaways & Limitations
The results offer a principled stabilization approach for adversarial training where standard gradient descent-ascent methods fail.
Source: §7 · p. 8The framework uses fixed control parameters for local stability guarantees, and discrete-time stability of iterative updates remains to be established.
Source: §4.2 · p. 5, §7 · p. 8
Abstract
from arXiv · showhide
Minimax optimization is fundamental to adversarial learning yet suffers from severe stability challenges, often manifesting as persistent limit cycles or divergence. This paper investigates these instabilities through the lens of continuous-time dynamical systems. By modeling training dynamics as a second-order Ordinary Differential Equation (ODE), we rigorously identify the root cause of non-convergence: the problematic eigenstructure of the loss function's Hessian near equilibrium points. To address this, we introduce Controlled-Minimax, a novel stabilization framework that reformulates the optimization trajectory as a controllable physical system. Unlike heuristic regularization techniques, our approach derives a specific feedback controller that provides theoretical guarantees of local exponential stability, regardless of the underlying Hessian spectrum. Consequently, our method stabilizes general minimax problems where traditional gradient descent-ascent methods fail. Extensive experiments on pathological benchmarks including the Dirac-GAN, 2D Gaussian mixtures, and Swiss Roll datasets, as well as high-resolution image synthesis with StyleGAN2-ada, demonstrate that Controlled-Minimax effectively suppresses oscillatory behavior while ensuring fast and robust convergence.
1. Introduction
Minimax training can oscillate or diverge, while existing guarantees and stabilization methods have limitations in the non-convex, non-concave settings common in deep learning. Controlled-Minimax models training as a second-order ODE and adds feedback control, providing local stability guarantees and empirical suppression of oscillations across synthetic benchmarks and StyleGAN2-ada.
1. Introduction
- Deep-learning minimax landscapes are non-convex and non-concave, and Gradient Descent-Ascent can fail to converge, producing oscillations, limit cycles, or divergence.Prior guarantees often rely on restrictive conditions, while regularization can hamper expressivity and dynamic adjustments can add computational overhead.Source: §1 · pp. 1–2
- The method treats optimization as a controllable system, using a designed feedback controller to dissipate energy and guarantee local exponential stability regardless of Hessian structure.The training dynamics are modeled as a second-order ODE.Source: §1 · p. 2
- The paper identifies pathological Hessian eigenstructures as a root cause of non-convergence in standard minimax dynamics.Source: §1 · p. 2
- Controlled-Minimax suppresses oscillations and supports robust, rapid convergence across synthetic benchmarks and StyleGAN2-ada.On Dirac-GAN, the baseline cycles indefinitely while Controlled Minimax drives the parameters to the equilibrium.Source: §1 · p. 2, Figure 1 · p. 3
1.1. Related Work
- Prior convergence results rely on structured curvature or step-size conditions, while regularization and control variants retain restrictive assumptions or require complex modifications.Source: §1.1 · p. 2
2. Preliminary
The preliminary section establishes stability terminology for autonomous ODEs and relates local stability to the Jacobian spectrum at equilibrium. It distinguishes stable, asymptotically stable, exponentially stable, and marginally stable cases.
2. Preliminary
- An equilibrium is a point where the system's vector field vanishes; the section considers autonomous ODEs with locally Lipschitz dynamics.Source: §2 · p. 3
- Lyapunov stability bounds trajectories near equilibrium, asymptotic stability additionally requires convergence, and exponential stability specifies a local exponential rate.Source: §2 · p. 3
- Local linearization classifies stability using the Jacobian spectrum: strictly negative real parts imply local asymptotic stability, while any positive real part implies instability.Eigenvalues with nonpositive real parts and semisimple zero-real-part eigenvalues give marginal stability.Source: §2 · p. 3
3. Minimax Dynamics as a Second Order ODE
The section derives an uncontrolled second-order ODE for minimax training and analyzes local stability through the spectrum of an interaction matrix at equilibrium. Its classification distinguishes asymptotic, marginal, and unstable behavior by eigenvalue conditions.
3. Minimax Dynamics as a Second Order ODE
- Differentiating the continuous-time gradient flow under twice continuous differentiability of L yields an uncontrolled second-order ODE with Hessian-coupled accelerations.The paper presents this as the original dynamics before introducing a stabilizing damping controller.Source: §3 · p. 3
3.1. Local Stability Analysis
- At zero-velocity equilibria, the Jacobian’s lower-right interaction block M determines local stability.The equilibrium condition also requires the loss gradients to vanish.Source: §3.1 · p. 4
- Local stability is asymptotic when every eigenvalue of M has negative real part, marginal when all have nonpositive real parts and zero-real-part eigenvalues have matching algebraic and geometric multiplicities, and unstable if any has positive real part.Velocity perturbations decay exponentially in the asymptotically stable case, remain bounded in the marginal case, and grow exponentially in the unstable case.Source: §3.1 · p. 4
4. Controlled Dynamics Formulation
Controlled-Minimax adds feedback inputs to second-order minimax dynamics and analyzes the resulting closed-loop system near a stationary point. Its controller combines damping and cross-coupling, with a local exponential stability guarantee under stated regularity assumptions.
4. Controlled Dynamics Formulation
- The framework extends the natural second-order dynamics with control inputs and integrates the resulting controller into a practical optimization method.The controlled equations add uθ and uϕ to the parameter-velocity dynamics.Source: §4 · p. 4
- The analysis assumes a twice continuously differentiable objective with bounded Hessian blocks and a locally Lipschitz system, ensuring nearby solutions exist and are unique.The analysis is conducted near a stationary point satisfying first-order optimality conditions.Source: §4 · p. 4
4.1. Controller Design
- The linear velocity-feedback controller combines positive damping gains λθ and λϕ with a cross-coupling gain β and fixed matrix Ac.The feedback is designed to add damping and cross-coupling between the θ and ϕ dynamics.Source: §4.1 · p. 4
- Ac couples the leading r = min{dθ, dϕ} coordinates of θ and ϕ, leaving any remaining dimensions uncoupled.The matrix provides identity coupling on those leading coordinates.Source: §4.1 · p. 5
- Near equilibrium, linearized velocity dynamics use diagonal damping to dissipate energy and a skew-symmetric controller term to modify natural player coupling.The linearization is expressed as v̇ = Mv, with Hessian blocks evaluated at the equilibrium.Source: §4.1 · p. 5
4.2. Exponential Stability of the Controlled System
- The stability analysis gives a sufficient damping-gain condition robust to β, under which the full nonlinear closed-loop system is locally exponentially stable.The theorem’s guarantee assumes the regularity conditions stated for the controlled dynamics.Source: §4.2 · p. 5
- The theoretical guarantee assumes control inputs are available continuously and instantaneously, whereas implementation must compute them iteratively.The following algorithmic treatment addresses this discretization gap.Source: §4.2 · p. 5
5. The Controlled-Minimax Algoirthm
The Controlled-Minimax algorithm augments gradient descent-ascent with feedback control to stabilize minimax updates. On the Swiss Roll, its generated samples form a continuous, smooth spiral, unlike the baseline’s disjointed segments.
5. The Controlled-Minimax Algoirthm
- The method integrates the controller into standard gradient descent-ascent, shifting updates toward a trajectory that balances gradient exploitation with improved convergence stability.The procedure is presented as Controlled-Minimax Optimization.Source: §5 · pp. 5–6
- Each iteration computes both players’ gradients, forms control inputs from damping and cross-player gradient terms, then applies corrected descent-ascent updates.The algorithm uses learning rate η, damping gains λθ and λϕ, and coupling parameter β.Source: §5 · p. 5
- The controller combines dissipative damping with cross-coupled gradient correction to counter rotational dynamics and stabilize the trajectory toward equilibrium.The damping terms add artificial friction, while cross-coupling injects the opponent’s gradient information.Source: §5 · pp. 5–6
- On the Swiss Roll, Controlled-Minimax generates a continuous, smooth spiral matching the target distribution, while the standard GAN baseline produces disjointed segments.The figure presents qualitative results on the Swiss Roll dataset.Source: Figure 2 · p. 6
6. Experiments
Across synthetic benchmarks, Controlled-Minimax suppresses oscillations and recovers target distributions; on StyleGAN2-ada, it converges faster and achieves lower FID than the baseline across three datasets.
6.1. Dirac-GAN
- Controlled-Minimax rapidly drives the Dirac-GAN parameters to the origin, while the baseline cycles indefinitely without converging.The game’s unique Nash equilibrium is (θ, ψ) = (0, 0).Source: §6.1 · p. 6
6.2. 2D Gaussian
- Controlled-Minimax distributes probability mass across all eight Gaussian modes, while the Standard GAN baseline exhibits severe mode collapse.The target consists of eight distinct, separated Gaussian components.Source: §6.2 · p. 7, Figure 3 · p. 7
6.3. Swiss Roll
- Controlled-Minimax generates samples that follow the Swiss Roll spiral with uniform coverage, while the baseline struggles with the manifold and exhibits mode collapse.The generated samples also reproduce the target distribution’s noise characteristics.Source: §6.3 · p. 7
6.4. StyleGAN2-ada
- After 7 days, Controlled-Minimax lowers FID from 3.37 to 2.91 on CIFAR-10, 10.48 to 8.18 on LSUN Cat, and 7.81 to 5.46 on FFHQ.It also converges faster and shows reduced oscillation ranges across all three datasets.Source: §6 · p. 6, Table 1 · p. 8, §6.4 · p. 8
7. Conclusion and Discussion
The paper attributes minimax instability to the loss Hessian’s eigenstructure and proposes a control-theoretic feedback controller. It reports local exponential stability and robust empirical convergence, while identifying open work on adaptive control and discrete-time stability.
7. Conclusion and Discussion
- The framework’s feedback controller is designed to steer the optimization trajectory and address instability linked to the Hessian eigenstructure near equilibrium.The paper reports local exponential stability regardless of the underlying Hessian spectrum, including problems where standard gradient descent-ascent fails.Source: §7 · p. 8
- Across synthetic benchmarks and StyleGAN2-ada image synthesis, Controlled-Minimax suppresses oscillations and supports robust, rapid convergence.The authors characterize this as a principled solution to a persistent challenge in adversarial training.Source: §7 · p. 8
- Fixed control parameters guarantee local stability, while adaptive control for global guarantees and stability of discretized updates remain future directions.The paper calls for stability conditions directly on discrete-time update rules.Source: §7 · p. 8
A. Proof of Theorem 4.3
The proof uses a quadratic Lyapunov argument to establish exponential stability for the linear system, then transfers local exponential stability to the nonlinear system through its Jacobian at equilibrium.
A. Proof of Theorem 4.3
- The proof bounds the change in a quadratic Lyapunov function along linear-system trajectories by estimating diagonal and cross contributions, then combining those estimates in equation (14).Source: A. Proof of Theorem 4.3 · p. 11
- The resulting bound establishes exponential stability of v = 0 for the linear system.The proof uses the inequality 2ab ≤ a^2 + b^2 to bound the combined terms.Source: A. Proof of Theorem 4.3 · pp. 11–12
- For the nonlinear system, local Lipschitz continuity and a Jacobian matching the linear-system matrix at equilibrium yield local exponential stability.The proof invokes standard nonlinear stability theory.Source: A. Proof of Theorem 4.3 · p. 12